Each glowing point is one candidate reaction condition — a (temperature, catalyst loading) pair. Two competing objectives are evaluated for every point from smooth, chemically-motivated response functions:
Yield(T,C) = Ymax · exp(-((T-Topt)/σT)²) · (1-e^(-C/Ck)) · e^(-C/Cdecay)
E-factor(T,C) = a + b·C + c·(T/100)²
Domination: A ≻ B iff Yield(A) ≥ Yield(B) and E(A) ≤ E(B),
with at least one strict inequality.
Yield rises with catalyst loading up to a point (more active sites) then falls (catalyst-driven side reactions / poisoning), and peaks at an optimum temperature — too cold and the reaction is slow, too hot and selectivity drops. Waste (E-factor: kg waste per kg product) climbs steadily with both catalyst amount and temperature.
- Non-dominated sorting — every generation, points are ranked into fronts: rank 1 is the Pareto set (no other point beats it on both objectives at once), rank 2 is dominated by exactly the rank-1 points, and so on.
- Selection — parents are drawn preferentially from lower ranks (binary tournament by rank, tie-broken by objective spread), so the search pressure pushes the whole population toward the front.
- Crossover + mutation — a child's (T, C) is the parents' average plus Gaussian jitter sized by the mutation-strength slider, then clamped to the valid process window.
- Elitism — the current Pareto front is always carried into the next generation unchanged, so the front can only improve or hold steady, never regress.
This is the same idea real self-driving labs use when yield alone is not the whole story: a chemist rarely wants the single highest-yielding recipe if it also means ten times the solvent waste — the Pareto front is the menu of honest trade-offs between the two.